The Little Weather Balloon Company Taking on Google DeepMind

It’s been a wild few years in the typically tedious world of weather predictions. For decades, forecasts have been improving at a slow and steady pace — the standard metric is that every decade of development leads to a one-day improvement in lead time. So today, our four-day forecasts are about as accurate as a one-day forecast was 30 years ago. Whoop-de-do.

Now thanks to advances in (you guessed it) artificial intelligence, things are moving much more rapidly. AI-based weather models from tech giants such as Google DeepMind, Huawei, and Nvidia are now consistently beating the standard physics-based models for the first time. And it’s not just the big names getting into the game — earlier this year, the 27-person team at Palo Alto-based startup Windborne one-upped DeepMind to become the world’s most accurate weather forecaster.

“What we’ve seen for some metrics is just the deployment of an AI-based emulator can gain us a day in lead time relative to traditional models,” Daryl Kleist, who works on weather model development at the National Oceanic and Atmospheric Administration, told me. That is, today’s two-day forecast could be as accurate as last year’s one-day forecast.

All weather models start by taking in data about current weather conditions. But from there, how they make predictions varies wildly. Traditional weather models like the ones NOAA and the European Centre for Medium-Range Weather Forecasts use rely on complex atmospheric equations based on the laws of physics to predict future weather patterns. AI models, on the other hand, are trained on decades of prior weather data, using the past to predict what will come next.

Kleist told me he certainly saw AI-based weather forecasting coming, but the speed at which it’s arriving and the degree to which these models are improving has been head-spinning. “There's papers coming out in preprints almost on a bi-weekly basis. And the amount of skill they've been able to gain by fine tuning these things and taking it a step further has been shocking, frankly,” he told me.

So what changed? As the world has seen with the advent of large language models like ChatGPT, AI architecture has gotten much more powerful, period. The weather models themselves are also in a cycle of continuous improvement — as more open source weather data becomes available, models can be retrained. Plus, the cost of computing power has come way down, making it possible for a small company like Windborne to train its industry-leading model.

Founded by a team of Stanford students and graduates in 2019, Windborne used off-the-shelf Nvidia gaming GPUs to train its AI model, called WeatherMesh — something the company’s CEO and co-founder, John Dean, told me wouldn’t have been possible five years ago. The company also operates its own fleet of advanced weather balloons, which gather data from traditionally difficult-to-access areas.

Standard weather balloons without onboard navigation typically ascend too high, overinflate, and pop within a matter of hours (thus becoming environmental waste, sad!). Since it’s expensive to do launches at sea or in areas without much infrastructure, there’s vast expanses of the globe where most balloons aren’t gathering any data at all.

Satellites can help, of course. But because they’re so far away, they can’t provide the same degree of fidelity. With modern electronics, though, Windborne found it could create a balloon that autonomously changes altitude and navigates to its intended target by venting gas to descend and dropping ballast to ascend.

“We basically took a lot of the innovations that lead to smartphones, global satellite communications, all of the last 20 years of progress in consumer electronics and other things and applied that to balloons,” Dean told me. In the past, the electronics needed to control Windborne’s system would have been too heavy — the balloon wouldn’t have gotten off the ground. But with today’s tiny tech, they can stay aloft for up to 40 days. Eventually, the company aims to recover and reuse at least 80% of its balloons.

The longer airtime allows Windborne to do more with less. While globally there are more than 1,000 conventional weather balloons launched every day, Dean told me, “We collect roughly on the order of 10% or 20% of the data that NOAA collects every day with only 100 launches per month.” In fact, NOAA is a customer of the startup — Windborne already makes millions in revenue selling its weather balloon data to various government agencies.

Now, with a potentially historic hurricane season ramping up, Windborne has the potential to provide the most accurate data on when and where a storm will touch down.

Earlier this year, the company used WeatherMesh to run a case study on Hurricane Ian, the Category 5 storm that hit Florida in September 2022, leading to over 150 fatalities and $112 billion in damages. Using only weather data that was publicly available at the time, the company looked at how accurately its model (had it existed back then) would have tracked the hurricane.

Very accurately, it turns out. Windborne’s predictions aligned neatly with the storm’s actual path, while the National Weather Service’s model was off by hundreds of kilometers. That impressed Khosla Ventures, which led the company’s $15 million Series A funding round earlier this month. “We haven’t seen meaningful innovation in weather since The Weather Channel in the 90s. Yet it’s a $100 billion market that touches essentially every industry,” Sven Strohband, a partner and managing director at Khosla Ventures, told me via email.

With this new funding, Windborne is scaling up its fleet of balloons as it prepares to commercialize. The money will also help Windborne advance its forecasting model, though Dean told me robust data collection is ultimately what will set the company apart. “In any kind of AI industry, whoever has the top benchmark at any given time, it’s going to fluctuate,” Dean said. “What matters is the model plus the unique datasets.”

Unlike Windborne, the tech giants with AI-based weather models — including, most recently, Microsoft — aren’t gathering their own data, instead drawing solely on publicly accessible information from legacy weather agencies.

But these agencies are starting to get into the game, too. The European Centre for Medium-Range Weather Forecasts has already created its own AI-based model, the Artificial Intelligence/Integrated Forecasting System, which it runs in parallel to its traditional model. NOAA, while a bit behind, is also looking to follow suit.

“In the end, we know we can't rely on these big tech companies to just keep developing stuff in good faith to give to us for free,” Kleist told me. Right now, many of the top AI-based weather models are open source. But who knows if that will last? “It's our mission to save lives and property. And we have to figure out how to do some of this development and operationalize it from our side, ourselves,” Kleist said, explaining that NOAA is currently prototyping some of its own AI-based models.

All of these agencies are in the early stages of AI modeling, which is why you likely haven’t noticed weather predictions making a pronounced leap in accuracy as of late. It’s all still considered quite experimental. “Physical models, the pro is we know the underlying assumptions we make. We understand them. We have decades of history of developing them and using them in operational settings,” Kleist told me. AI-based models are much more of a black box, and there’s questions surrounding how well they will perform when it comes to predicting rare weather events, for which there might be little to no historical data for the model to reference.

That hesitation might not last long, though. “To me it’s fairly obvious that most of the forecasts that would actually be used by users in the future will come from machine learning models,” Peter Dueben, head of Earth systems modeling at the European Centre for Medium Range Weather Forecasting, told me. “If you just want to get the weather forecast for the temperature in California tomorrow, then the machine learning model is typically the better choice,” he added.

That increased accuracy is going to matter a lot, not just for the average weather watcher, but also for specific industries and interest groups for whom precise predictions are paramount. “We can tailor the actual models to particular sectors, whether it's agriculture, energy, transportation,” Kleist told me, “and come up with information that's going to be at a very granular, specific level to a particular interest.” Think grid operators or renewable power generators who need to forecast demand or farmers trying to figure out the best time to irrigate their fields or harvest crops.

A major (and perhaps surprising) reason this type of customization is so easy is because once AI-based weather models are trained, they’re actually orders of magnitude cheaper and less computationally intensive to run than traditional models. All of this means, Kleist told me, that AI-based weather models are “going to be fundamentally foundational for what we do in the future, and will open up avenues to things we couldn't have imagined using our current physical-based modeling.”